Technical Field
The present invention relates to estimating abnormal sensors.
Related Art
A conventional method includes mounting a plurality of sensors in a complex system such as an automobile or manufacturing apparatus, analyzing a plurality of pieces of time sequence data acquired from the plurality of sensors, and monitoring the presence of abnormalities, as shown in Japanese Patent Application Publication No. 2010-78467, for example.
With such a method, the amount of sensor abnormality is scored and identified based on the amount of change in the relational structure between groups of sensors including a group of sensors indicating detection results for which an input signal is within a normal signal range, i.e. normal sensors, and a group of sensors indicating detection results for which the input signal is in an abnormal signal range, i.e. abnormal sensors, from the pieces of time sequence data. However, with a method such as described in Japanese Patent Application Publication No. 2010-78467, the relational structure between the sensors has a linear relationship, and therefore it is difficult to handle a complex non-linear relationship that actually occurs among a plurality of sensors. In other words, with the conventional processing method, even when there is normal data corresponding to a non-linear relationship among the sensors, this data could be judged to be abnormal due to being an outlier from a linear model. Furthermore, with the conventional processing method, even when there is abnormal data that is an outlier from the non-linear relationship among the sensors, this data could be judge to be normal if the data matches a linear model.